MEMS vibration sensor-based edge AI for machinery fault prediction: feasibility study using a petrochemical plant process simulation facility

Because of the presence of large quantities of flammable and explosive substances, a petrochemical plant requires artificial intelligence (AI)-based monitoring systems to enhance safety and mitigate accident risks. Herein, we demonstrate the feasibility of using microelectromechanical system (MEMS) vibration sensors in petrochemical plants by experimentally comparing their performance with those of conventional vibration sensors on the basis of the prediction accuracy of a one-dimensional time-series convolutional neural network model. In particular, we established a petrochemical plant process simulation facility to effectively collect anomaly data, which is exceptionally rare in real-world petrochemical plants. The petrochemical plant process simulation facility was employed to simulate fixture looseness, and two types of leak conditions as well as normal operation. Then, a MEMS sensor was used to collect six-axis data from both its accelerometer and gyroscope, while a conventional sensor captured only three-axis data from its accelerometer. When considering single-axis data, the MEMS sensor demonstrated superior classification accuracy (85.46%) compared to the conventional vibration sensor (80.94%). Moreover, when multiaxis data were used, with six and three axes from the MEMS and conventional sensors, respectively, both systems achieved similar performance outcomes (MEMS sensor: 99.91%, conventional sensor: 99.94%). These results indicate that MEMS sensors can effectively complement conventional vibration sensors, offering a cost-effective and scalable approach for monitoring petrochemical plants.